BlogVulnerability Research

GPUStack (CVE-2026-102895): Unauthenticated Admin on the Inference Layer

CVE-2026-102895 is a CVSS 9.8 flaw in GPUStack — the popular self-hosted GPU-and-model-management platform — that lets an unauthenticated attacker take over an admin account and disclose the API keys and cloud secrets the control plane holds. CISA added it to the Known Exploited Vulnerabilities catalog on 9 November 2026. If you run self-hosted inference and your GPUStack control plane has ever been reachable from outside a tight private network, treat it as compromised until it is patched, isolated, and every key it held is rotated. No exploitation details below.

6 min readBy TigerGate Research, Security Research TeamPublished Updated

TL;DR

  • CVE: CVE-2026-102895. CVSS 9.8 Critical. Unauthenticated admin takeover on GPUStack.
  • Downstream exposure: API keys and cloud secrets stored in the control plane — LLM-provider keys, GPU-host credentials, object-storage access.
  • CISA KEV: added 9 November 2026 — 21-day remediation clock for federal civilian agencies; many private contracts reference the same.
  • Fix: patch to the vendor's fixed release; confirm the exact version from the GPUStack advisory when you apply.
  • Posture question: is the GPUStack admin interface reachable from outside a tight private network? If yes, treat as compromised until patched + rotated.

What's affected

GPUStack is the self-hosted platform teams use to schedule GPU hosts, manage model deployments, and front downstream inference infrastructure. CVE-2026-102895 sits in the authentication layer of the control plane: an unauthenticated attacker can reach admin, enumerate stored credentials, and pivot. The blast radius matches what an inference control plane typically holds:

  • Hosted model-provider API keys (OpenAI, Anthropic, Google, AWS Bedrock, Azure OpenAI).
  • Cloud credentials for the GPU hosts GPUStack schedules against.
  • Object-storage credentials for model weights and training artefacts.
  • Any internal integration token the instance was configured with.

"Posture-first" is the right frame here: the CVE is bad, but the deployment pattern — admin interface reachable on 0.0.0.0, no upstream auth gate, default credentials never rotated — is what turns it into a full-cloud compromise.

What to do

  1. Patch to the vendor's fixed release; confirm the exact fixed version from GPUStack's advisory when you apply.
  2. Take the control plane off the public internet. Bind to private interfaces, front with authenticated ingress, enforce IP allow-lists.
  3. Rotate every key and cloud credential the instance could have held — LLM-provider keys, cloud-host credentials, object-storage access, integration tokens. Treat the control plane as compromised until proven otherwise if it was ever reachable from outside a tight private network.
  4. Review admin and API access logs for unauthenticated calls or new admin activity in the window leading up to patching.
  5. Inventory your model-serving surface. Anywhere you run GPUStack, Ray, vLLM, Ollama, LocalAI, OpenWebUI, or similar self-hosted inference tooling. Each is a control-plane-posture question even before the next CVE.

A short LiteLLM note (same audience)

Not on KEV as of 17 November (re-check on your patch day): CVE-2026-114612, CVSS 9.9, in LiteLLM. In affected versions, secret storage and session-token signing share a single key, which lets an internal LiteLLM user forge a proxy_admin session and then execute code via a stdio MCP server. Fixed in LiteLLM 1.100.4, 1.101.3, 1.102.2 and 1.103.1. If you operate a LiteLLM gateway, upgrade to one of those versions, rotate your signing / secret key material as part of the upgrade, and inventory any internal users with reachability to the admin surface.

For broader LLM-gateway guidance see our LiteLLM RCE and LLMjacking economics post, and the how-to-stop-LLMjacking controls checklist.

Companion advisory — same patch window

SSHole (CVE-2026-123158) also landed on CISA KEV on 9 November. It is an OpenSSH authentication bypass on musl / Alpine systems, exploited within 40 hours of the fix. If you run containers and self-hosted inference — most teams do — one patch window should cover both.

How TigerGate helps

  • AI-SPM inventories self-hosted inference control planes (GPUStack, Ray, vLLM, Ollama, LocalAI, OpenWebUI), flags admin surfaces reachable from outside a private network, and tracks the API-key and cloud-credential exposure held by each instance.
  • CSPM catches the network side: GPUStack and LLM-gateway ports reachable from 0.0.0.0/0 in AWS security groups, Azure NSGs, GCP firewall rules, and Kubernetes NetworkPolicies.
  • Identity Security (CIEM) surfaces the AI-provider and cloud identities held by GPUStack and alerts on anomalies: idle credentials suddenly reactivating, roles accessing secrets they never previously touched.
  • eBPF runtime detection watches for anomalous admin activity on GPUStack and similar control planes, and anomalous outbound to AI-provider endpoints from inference hosts.
  • Compliance evidence maps patch and rotation activity to SOC 2 CC7.1, ISO 27001 A.8.8, PCI DSS 6.3.3, and the BOD 22-01 clock on CISA KEV entries.

FAQ

CVE-2026-102895 (CVSS 9.8) allows an unauthenticated attacker to take over a GPUStack admin account. Once admin, the attacker can enumerate and disclose stored API keys and cloud secrets — the credentials GPUStack holds to orchestrate GPU hosts and talk to downstream LLM providers and object stores. It is on CISA's KEV catalog as of 9 November 2026.
Any team running self-hosted model-serving on GPUStack, especially where the admin interface or API is reachable beyond a tight private network. Confirm the fixed release from GPUStack's vendor advisory when you apply the patch.
Take the control plane off the public internet, rotate every key and cloud credential the instance could have held, and review admin and API access logs for unauthenticated calls or new admin activity in the window leading up to patching. Treat the control plane as compromised until proven otherwise if it was ever reachable from outside a tight private network.
CVE-2026-114612 (CVSS 9.9) in LiteLLM lets an internal user forge a proxy_admin session because secret storage and session signing share a key, then execute code via a stdio MCP server. Not on KEV as of publication — re-check on your patch day. Fixed in LiteLLM 1.100.4, 1.101.3, 1.102.2 and 1.103.1.

Sources

Scan your inference surface with TigerGate

14-day trial. TigerGate inventories every self-hosted inference control plane, flags unpatched GPUStack and LiteLLM, and maps the AI-provider / cloud credentials each instance holds — so one patch window closes the whole AI stack.